DOI: 10.64808/engineeringperspective.1932616 ISSN: 2757-9077

Performance Analysis of Modern and Traditional Convolutional Neural Networks in the Classification of Optical Coherence Tomography (OCT) Images

Kıyas Kayaalp, İlkay Onay
Show PDF Cite
Optical Coherence Tomography (OCT) plays a critical role in the early diagnosis of retinal diseases. In this study, a deep learning-based automatic diagnosis system was developed using the OCTMNIST dataset, which includes Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), Drusen, and normal retina images. This study was conducted to measure the capability of artificial intelligence in this field. To increase rigorous evaluation and prevent data leakage, the dataset was separated into isolated folder structures: 80% training, 10% validation, and 10% testing. Within the scope of the study, five different models consisting of traditional and modern Convolutional Neural Network (CNN) architectures (ResNet18, MobileNetV2, EfficientNetB0, DenseNet121, and VGG16) were trained and their performances were compared. During the training process, Automatic Mixed Precision (AMP) was used for hardware optimization, and Early Stopping algorithms were utilized to prevent overfitting. According to the test results, the EfficientNetB0 model achieved the highest success with 93.67% accuracy and a 0.88 F1-Score. In contrast, older traditional architectures like VGG16 required architecture-specific hyperparameter tuning just to converge, highlighting the superior stability of modern CNNs. The research results demonstrate that modern lightweight CNN architectures offer high efficiency in medical image processing. Furthermore, to provide a practical approach for clinical use, the most successful model was integrated into a web-based interface and presented to the end-user. This study reveals the performance differences of deep learning models under hardware constraints and proposes an end-to-end artificial intelligence solution that will accelerate medical diagnosis processes.

More from our Archive